commvita uses AI where it saves time and never where it would take a clinical decision. There is one governed configuration (provider keys server-side only), AI at the point of care that suggests but never decides, executive decision-support with a mandatory trust envelope — and a governance spine (ATRS, DCB0129/0160, MHRA AIaMD, GDPR Art.22) over all of it. Non-SaMD throughout.
Configured once, used safely, evidenced always. No patient data leaves for a model without a server-side key and an audit entry; every AI surface is non-SaMD — it provides context and suggestions, a human makes the call.
The single, canonical AI configuration — Claude / Microsoft Copilot / OpenAI, one active provider. API keys are server-side environment only (never in the browser). The Ask commvita assistant and every AI feature call the same governed proxy.
Ambient Clinical AI turns a consultation into a structured SOAP note with SNOMED auto-coding (audio never persisted); VoiceScribe, AI Consultation Assistant and NLP Document Coding assist documentation. Every code is accept/reject — the clinician decides.
The AI Transformation Advisor turns a strategic objective + constraints into options, prioritised interventions, KPIs and a roadmap — each with a mandatory trust envelope (assumptions, confidence, human-review areas, evidence) and a 3-way scenario comparison. Gated to executive + admin.
ANTHROPIC_API_KEY etc.), never entered in or returned to the browser.| Provider | Model / endpoint | Key | Active |
|---|---|---|---|
| Claude (Anthropic) | claude-sonnet-4-6 | present | ● active |
| Microsoft Copilot (Azure OpenAI) | gpt-4o · uksouth | present | — |
| OpenAI | gpt-4o | not set | — |
Admins pick the active provider; the key comes from a server-side env var (mirrors the SMTP pattern). Test-connection checks key presence only. The assistant and all features read this one config.
Point-of-care AI is non-SaMD: it drafts notes and proposes codes, and the clinician accepts, edits or rejects. Audio is transcript-only; the record is the clinician's.
The AI Governance module holds an ATRS model registry, DCB0129/0160 clinical-safety, NICE ESF tiers, MHRA AIaMD surveillance and a GDPR Art.22 override log; bias monitoring flags any disparity ratio outside 0.8–1.2 for DPO review.
The AI Governance module is where “is it safe and is it governed?” is answered with artefacts, not assurances: every tool carries a published model card (ATRS) and a NICE ESF risk tier; fairness is monitored quarterly against the Equality Act with a hard ±20% disparity threshold; and every deployment has an Algorithmic Impact Assessment covering DCB0129/0160, MHRA post-market surveillance and the devolved-nation gateways (Scotland PBPP, Wales HTW) plus the GDPR Art.22 right to human review.
| AI tool | NICE ESF tier | Clinical safety | ATRS |
|---|---|---|---|
| Ambient Clinical AI | Tier C — Inform | DCB0129 ✓ | published |
| Deterioration ML (continuous RPM) | Tier C — Drive | DCB0129/0160 ✓ | published |
| Transformation Advisor | Tier A | n/a (non-clinical) | published |
| Ask commvita assistant | Tier A | n/a (non-clinical) | published |
| Subgroup | Metric | Value | Reference | Disparity | Status |
|---|---|---|---|---|---|
| Age 75+ | Sensitivity | 0.86 | 0.90 | 0.96 | within |
| Female | Sensitivity | 0.88 | 0.90 | 0.98 | within |
| Ethnicity — Black | PPV | 0.71 | 0.90 | 0.79 | review |
| IMD decile 1 | Recall | 0.84 | 0.90 | 0.93 | within |
| Capability | Route | Model / API | Status |
|---|---|---|---|
| AI Integration Hub (single config) | /ai-integration-hub | API /ai/config · /ai/active-provider | ● Live |
| Ask commvita assistant (chat proxy) | global overlay | API /ai/assistant/chat · keyword fallback | ● Live |
| AI Transformation Advisor | /ai-transformation-advisor | API /transformation/cases · trust envelope | ● Live |
| Ambient Clinical AI (voice → SOAP) | /ambient-clinical-ai | SNOMED auto-code · hazard log | ☉ Demonstrated |
| VoiceScribe · AI Consult Assist | /voice-scribe · /ai-consult-assist | Web Speech · suggestion panel | ☉ Demonstrated |
| NLP Document Coding | /nlp-coding | SNOMED · confidence · review queue | ☉ Demonstrated |
| AI Governance (ATRS · bias) | /ai-governance | ATRS · DCB0129/0160 · Art.22 | ☉ Demonstrated |
| ML: continuous RPM · A&E demand · LOS | /continuous-rpm · /ae-demand-forecast | decision-support (non-SaMD) | ☉ Demonstrated |